Deep Research is one of the most ambitious features Google has built for Gemini. Instead of answering a question from training data, Deep Research goes out and investigates – searching the web, reading multiple sources, following leads, and synthesizing everything into a structured report with citations.

The difference between asking Gemini a question and asking it to deep-research a question is the difference between getting an answer and getting an investigation. An answer is one model’s best guess from training. An investigation finds current information, cross-references it, and presents verifiable findings.

Google has made the Deep Research API publicly available at ai.google.dev/gemini-api/docs/deep-research. The technology works. The question is how well it works on Mac, and what is missing.

TL;DR – Deep Research at a glance

What it is: a four-step agent that plans, searches, reads, and synthesizes a multi-source report on any topic you give it
Typical runtime: 5-15 minutes per query, executing 10-20 web searches and reading full pages for each
Output: a structured Markdown report with executive summary, topical sections, and inline citations linking back to sources
Model: gemini-2.5-flash-preview-04-17 with the google_search tool enabled (distinct from standard chat)
Availability: Gemini Advanced subscribers on web, Android, and iOS. Listed as a feature on the native Mac app at launch; rollout is per-account
Rate limits: Advanced caps Deep Research at a few dozen runs per day and queues requests under load
Report size: typically 2,500-6,000 words with 15-40 citations
Best on desktop because: reports are documents, not chat turns – you need a large screen to read them, a file system to save them, and desktop tools (Obsidian, VS Code, Notion) to integrate them

Windows note: Google has no native Gemini Windows app. Deep Research on Windows is a browser-only experience on gemini.google.com unless you install a third-party native client. The native Gemini Windows app guide walks through the options, and Google didn’t make a Gemini Windows app explains why the Windows “Google app for desktop” does not count.

How Deep Research works

Deep Research follows a four-step agentic workflow.

Step 1: Research plan. Gemini generates a structured outline – sub-questions to explore, source types to seek, report structure. The plan is visible before execution. You can add topics, remove tangents, or redirect focus before Gemini spends compute on searches.

Step 2: Multi-step web search. Gemini executes multiple searches with different query formulations targeting different aspects of the plan. If initial results are insufficient, it reformulates and searches again. A human researcher might run three or four searches before tiring. Deep Research routinely executes ten to twenty per investigation.

Step 3: Source reading. For each result, Deep Research reads the full page – not just the search snippet. It extracts relevant information, notes source URLs for citation, and identifies claims needing cross-referencing. Reading ten to twenty full pages and holding all of it in context takes a human hours. Deep Research does it in minutes.

Step 4: Synthesis. Findings are synthesized into a structured report: executive summary, detailed sections by topic, inline citations with source URLs, areas of consensus and disagreement, and gaps in available information.

For a longer glossary definition and a comparison to how “deep research” is marketed across vendors (Perplexity Deep Research, OpenAI’s Deep Research mode, Claude Projects with web search), see our Deep Research glossary entry.

Step-by-step: running a Deep Research query on Mac

Open Gemini for Mac (Option+Shift+Space opens the full window, which is what you want for Deep Research – the mini window truncates report output).
Tap the Deep Research chip above the compose bar. If the chip is missing, you are on a free tier – Deep Research requires Gemini Advanced.
Type your research question. Favor specificity: “Compare the revenue and margin profiles of Cloudflare, Akamai, and Fastly for FY2023 and FY2024” beats “compare CDN companies.”
Review the research plan. Gemini presents a bulleted outline of sub-questions and sources before it spends compute. Edit freely: add angles, cut tangents, specify preferred source types (academic, industry reports, first-party docs).
Hit Start research. Expect 5-15 minutes. A progress strip at the top shows current step and rough ETA.
Read the report in the full window. Citations are inline and clickable; the sources panel lists every URL consulted.
Export. Click the … menu above the report. Options typically include “Copy as Markdown,” “Export to Google Docs,” and in the latest Mac builds “Save as .md.” Use Markdown for local knowledge bases, Docs for team sharing.
Follow up. Ask Gemini to expand a section, add a new sub-topic, or re-run with different sources. Each follow-up is a new Deep Research turn that references the original.

One note on plans: on the first day of the Mac launch, a subset of users saw Deep Research fall back to single-turn grounded search instead of the full multi-step agent. If your “report” comes back as a short answer with five citations, it was not the real Deep Research – open the … menu and verify the model used says gemini-2.5-flash-preview-04-17 + google_search. If it says anything else, cancel and re-run in a fresh conversation.

Status on the official Mac app

Google’s landing page for Gemini for Mac lists Deep Research as a feature. The web interface supports it for Gemini Advanced subscribers. However, the degree to which Deep Research is fully operational in the native Mac app at launch is unclear.

The API documentation shows Deep Research uses the gemini-2.5-flash-preview-04-17 model with a specific google_search tool configuration – it is a distinct capability from standard chat, not simply “a harder question.”

As of launch (April 15, 2026), the most common user-reported gaps in the native Mac app versus the web experience are:

No persistent task queue. You cannot line up three research tasks, walk away, and come back to all three completed. Each report runs in its active conversation and blocks further activity in that conversation until done.
No export-to-file by default. Reports live in chat history. You copy Markdown manually. Obsidian vault integration is not built in.
Rate-limit visibility is thin. When you hit Advanced’s daily Deep Research ceiling, the UI says “try again later” without a timer or clear quota readout.
Canvas handoff is ad hoc. You can ask Gemini to “turn this report into a one-page summary infographic,” which triggers Canvas, but the research citations are not automatically preserved in the Canvas output.

Why Deep Research matters more on desktop

Deep Research produces long-form output – reports running several thousand words with dozens of citations. This format is inherently a desktop use case.

Reading and reviewing. A 3,000-word report with inline citations demands a large screen and the ability to have the report and sources open side by side. Reading it on a phone means constant scrolling and tiny citation links.

Saving and organizing. Research reports are reference documents you produce once and revisit. On desktop, a report saved as Markdown in your Obsidian vault becomes part of your knowledge base. In a mobile chat history, it is effectively ephemeral.

Iterating. Reports prompt follow-up questions. You identify a claim needing more detail and research that aspect further. Desktop lets you have the original report open in one window while conducting follow-up research in another.

Sharing. Reports need to reach colleagues and collaborators via email, Slack, or shared drives. A report in a chat thread is difficult to share with anyone outside your Gemini account.

Deep Research vs alternatives on Mac

Tool	Model	Pricing	Export	Desktop-first?
Gemini Deep Research (Mac)	Gemini 2.5 Flash + google_search	Included with Advanced	Chat history, Markdown copy, Docs	DMG app, no file integration
Perplexity Deep Research	Sonar + reasoning model	Free tier + Pro	Markdown, PDF, share link	Web + Mac app
ChatGPT Deep Research	o-series reasoning	Plus / Pro	Chat history, Docs, Markdown	Mac app
Claude Projects + web search	Claude Sonnet 4	Pro	Chat + Project artifacts	Claude Desktop
GeminiDesktop.app + API	Gemini 2.5 via API	API-priced per run	Auto-save to local .md, Obsidian-ready	Native Mac, Intel + Silicon

The pattern: every vendor has a “deep research” mode now, and all of them produce comparable-quality first drafts. The differentiator on Mac is no longer “which model,” it is “how fast can I turn the output into a document I actually own.” That is why local file export and Obsidian-readiness matter more than raw model benchmarks.

How GeminiDesktop approaches Deep Research

We are building GeminiDesktop with Deep Research as a first-class desktop feature.

Direct API integration

GeminiDesktop connects directly to the Gemini Deep Research API. Submit a research query and the full pipeline – plan, search, reading, synthesis – runs with progress visible in the interface. You review the plan before execution and receive the final report with full citation metadata.

Reports saved as local Markdown

Every report is automatically saved as a Markdown file on your local file system – full text, citations as Markdown links, metadata, and the original query. You choose where files are saved: Documents folder, project directory, Obsidian vault.

This is a fundamental difference from browser-based Deep Research, where reports live in chat history and require manual copy-paste to save. In GeminiDesktop, every report is a first-class document from the moment it is generated.

Task queue

Queue multiple research questions. GeminiDesktop executes them sequentially (or in parallel, subject to your API quota), so you can kick off “Q1 competitive landscape,” “regulatory changes in EU AI Act,” and “internal team hiring trends in GenAI” before lunch and come back to three saved reports after.

Offline access

Local Markdown files mean you can read reports without an internet connection. Your research library does not depend on Gemini’s servers. Research conducted yesterday is available on a flight today.

Knowledge tool integration

Local Markdown integrates with tools researchers already use. Obsidian indexes and searches them across your vault. VS Code renders them with syntax highlighting. Any Markdown editor opens them. Reports become part of your existing knowledge infrastructure, not a separate silo.

The desktop research workflow

Type a research question in GeminiDesktop.
Review and approve the research plan.
Watch the multi-step search and synthesis execute.
Read the formatted report with citations.
The report saves automatically as Markdown in your chosen directory.
Open it in Obsidian, link it to related notes, build your knowledge base.
Ask follow-up questions – new reports reference the original and save alongside it.

Every step happens on your desktop. The reports are yours – local files you control and organize.

Troubleshooting common Deep Research issues on Mac

“Deep Research chip is missing from compose bar.” You are signed in on a free Gemini account, or your Advanced subscription hasn’t propagated. Sign out, sign back in, and verify your plan at gemini.google.com/app/plan.

Research stalls at “Reading source 7 of 12.” Usually a page Gemini is fetching has a slow TTFB or blocks headless fetches. The pipeline moves on after ~30 seconds. If it hangs longer than 2 minutes, cancel and retry. Narrow the research plan to fewer sources.

Report is oddly short (500 words, 3 citations). You hit the grounded-search fallback instead of the full Deep Research agent, typically because of high load or rate limiting. Start a fresh conversation and retry with the Deep Research chip explicitly toggled.

Citations link to paywalled PDFs. Gemini sees the abstract or the Google-cached snippet, not the full PDF. The citation is honest about what Gemini actually read. For paywalled content, verify manually.

“You’ve reached your Deep Research limit for today.” Gemini Advanced caps Deep Research at a few dozen runs per 24 hours. If you hit it, either wait (timer does not display, but quota resets rolling 24h from your first run of the day) or switch to the API via GeminiDesktop, which bills per-run instead of per-day.

Reports won’t save to Obsidian automatically. The native Mac app does not have Obsidian integration yet. Options: (a) export to Markdown manually and drop into your vault, (b) use GeminiDesktop’s local-save target configured to your vault, (c) run a small Hazel rule that watches your Downloads folder and moves new .md files into Obsidian.

Advanced tips for serious researchers

1. Pre-seed the research plan with citations you already trust

Before starting, paste 3-5 URLs of sources you know are authoritative for this topic. Gemini will include them in the plan and use them as anchors. This avoids the “Gemini found a content-marketing blog post and treated it as a peer” failure mode.

2. Use “restrict to last 24 months” for fast-moving topics

In your research prompt, add “Prioritize sources published in the last 24 months” for any topic where industry state is changing rapidly (AI, crypto, regulatory). Gemini respects this as a ranking signal, though not a hard filter.

3. Chain research -> Canvas -> export

Run Deep Research. Once you have the report, say “Turn the executive summary into a one-page infographic using Canvas.” You get the long-form document and a shareable visual, both linked to the same citations.

4. Keep a “research index” markdown file

Every time a Deep Research run saves a .md, append a one-line entry to a master research-index.md with the date, topic, and link to the file. After three months you have a searchable table of contents for your research library. This is trivial to automate with Hazel, Obsidian’s Dataview, or a Raycast script.

A 4,000-word report is rarely the right artifact to share with a stakeholder. Feed the report back to Gemini and ask for a 200-word executive briefing in bullet points. Share the briefing; link the full report as an appendix for anyone who wants the depth.

Closing the gap

Deep Research is one of Gemini’s most powerful capabilities and one of the most poorly served by a browser chat interface. Research reports are not chat messages. They are documents that need saving, organizing, referencing, and sharing. The browser treats them as ephemeral conversation turns. The desktop should treat them as persistent knowledge assets.

FAQ

Is Deep Research available on the free tier? No. Deep Research on the official Mac app requires Gemini Advanced. The Deep Research API is available to anyone on the Gemini API with a billing account.

Can I run Deep Research on multiple topics in parallel? Not in the native Mac app – each conversation blocks until its research completes. With the API (including via GeminiDesktop), parallel execution is limited only by your API quota.

Does Deep Research read behind paywalls? No. Deep Research sees whatever a logged-out public fetch would return – typically the abstract, summary, or Google-cached version. It will cite what it actually saw.

How fresh is the information? Deep Research pulls live web results at query time. A report on “latest iOS 18 accessibility features” will reflect whatever is on the web today, including articles published hours ago.

Can I direct Deep Research to specific domains? Yes, via your prompt. Say “Restrict sources to arxiv.org, nature.com, and pubmed.gov for this investigation.” Gemini respects domain constraints in its search formulation.

What happens to my research history? In the official Mac app, reports live in chat history attached to your Google account. They sync across devices via your Google account but do not save as local files by default. GeminiDesktop writes every report to disk and does not retain server-side history beyond what Google’s API requires.

We are building GeminiDesktop to be the client Deep Research deserves. Local-first output, Obsidian integration, offline access, and a research workflow designed for how people actually work. If you conduct research on your Mac, your reports should live on your Mac. Try GeminiDesktop.